Cathedral researchers reached some of Anthropic’s business systems by exploiting weaknesses in third-party software, The Washington Post reported on October 26. Anthropic and Cathedral gave different accounts of the scope and duration.
What could researchers reach?
A draft Cathedral report reviewed by the Post said its agents found routes to meeting transcripts, salary and job-offer information, and Slack discussions about advanced models in under 12 hours. Anthropic said researchers had first spent several days probing its own systems before turning to outside applications.
Anthropic confirmed access to a limited set of sales, human-resources and IT systems. These contained some customer and partner business contacts and basic user account records, it told the newspaper. The company said customer conversations, source code and model weights were not accessed.
A dispute over the testing rules
The work was conducted through a program that invites researchers to report security weaknesses. Anthropic said Cathedral had not fully disclosed its activity and had not followed all program rules. Cathedral’s draft said data access had been kept to a minimum. Those accounts remain distinct.
Anthropic said monitoring had flagged some activity before the report, that it cut off access the same day and disabled the affected business software until vendors addressed the weaknesses. The incident illustrates a separate boundary around connected business applications, even when the AI model itself was not accessed.
